Báo cáo khoa học: "Encoding a Parallel Corpus for Automatic Terminology" pot

Proceedings of EACL '99
Encoding a Parallel Corpus for Automatic Terminology
Extraction
Johann Gamper
European Academy Bolzano/Bozen
Weggensteinstr. 12/A, 39100 Bolzano/Bozen, Italy
j gamper@eurac, edu
Abstract
We present a status report about an
ongoing research project in the field
of (semi-)automatic terminology acquisi-
tion at the European Academy Bolzano.
The main focus will be on encoding a
text corpus, which serves as a basis for
applying term extraction programq.
1 Introduction
Text corpora are valuable resources in all areas
dealing with natural language processing in one
form or another. Terminology is one of these
fields, where researchers explore domain-specific
language material to investigate terminological is-
sues. The manual acquisition of terminological
data from text material is a very work-intensive
and error-prone task. Recent advances in auto-
matic corpus analysis favored a modern form of
terminology acquisition: (1) a corpus is a col-
lection of language material in machine-readable
form and (2) computer programs scan the cor-
pus for terminologically relevant information and
generate lists of term candidates which have to
be post-edited by humans. The following project

CATEx adopts this approach.
2
The CATEx Project
Due to the equal status of the Italian and the Ger-
man language in South Tyrol, legal and admin-
istrative documents have to be written in both
languages. A prerequisite for high quality trans-
lations is a consistent and comprehensive bilingual
terminology, which also forms the basis for an in-
dependent German legal language which reflects
the Italian legislation. The first systematic effort
in this direction was initiated a few years ago at
the European Academy Bolzano/Bozen with the
goal to compile an Italian/German legal and ad-
ministrative terminology for South Tyrol.
The CATEx (C_omputer A_.ssisted Terminology
E___~raction) project emerged from the need to sup-
port and improve, both qualitatively and quan-
titatively, the manual acquisition of terminologi-
cal data. Thus, the main objective of CATEx is
the development of a computational framework for
(semi-)antomatic terminology acquisition, which
consists of four modules: a parallel text corpus,
term-extraction programs, a term bank linked to
the text corpus, and a user-interface for browsing
the corpus and the term bank.
3 Building a
Parallel Text Corpus
Building the text corpus comprises the following
tasks: corpus design, preprocessing, encoding pri-

mary data, and encoding linguistic information.
3.1 Corpus Design and Preprocessing
Corpus design
selects a collection of texts which
should be included in the corpus. An important
criteria is that the texts represent a realistic model
of the language to be studied (Bowker, 1996). In
its current form, our corpus contains only one sort
of texts, namely the bilingual version of Italian
laws such as the Civil Code. A particular feature
of our corpus, which contains both German and
Italian translations, is the structural equivalence
of the original text and its translation down to the
sentence level, i.e. each sentence in the original
text has a corresponding one in the translation.
The corpus is one of the largest special language
corpora. It contains ca. 5 Mio. words and 35,898
(66,934) different Italian (German) word forms.
In the
preprocessing
phase we correct (mainly
OCR) errors in the raw text material and produce
a unified electronic version in such a way as to
simplify the programs for consequent annotation.
3.2 Encoding Primary Data and
Linguistic Annotation
Corpus encoding successively enriches the raw
text material with explicitly encoded informa-
275
Proceedings of EACL '99

tion. We apply the Corpus Encoding Standard
(CES), which is an application of SGML and pro-
vides guidelines for encoding corpora that are used
in language engineering applications (Ide et al.,
1996). CES distinguishes primary data (raw text
material)
and
linguistic annotation (information
resulting from linguistic analyses of the raw texts).
Primary data encoding covers the markup of
relevant objects in the raw text material. It com-
prises documentation information (bibliographic
information, etc.) and structural information
(sections, lists, footnotes, references, etc.). These
pieces of information are required to automati-
cally extract the source of terms, e.g. "Codice
Civile, art. 12". Structural information helps also
to browse the corpus; this is important in our case,
since the corpus will be linked to the terminolog-
ical database.
Encoding linguistic annotation enriches the pri-
mary data with information which results from
linguistic analyses of these data. We consider the
segmentation of texts into sentences and words,
the assignment/disambiguation of lemmas and
part-of-speech (POS) tags, and word alignment.
Due to the structural equivalence of our paral-
lel texts, we can easily build a perfectly sentence-
aligned corpus which is useful for word alignment.
The above mentioned linguistic information is re-

quired for term extraction, which is mainly in-
spired by the work in (Dagan and Church, 1997).
The monolingual recognition of terms is based on
POS patterns which characterize valid terms and
the recognition of translation equivalents is based
on bilingual word alignment. Lemmas abstract
from singular/plural variations, which is useful for
alignment and term recognition.
4 Discussion
The general approach we adopted in the prepro-
cessing and primary data encoding phases was
to pass the raw texts through a sequence of fil-
ters. Each filter adds some small pieces of new
information and writes a logfile in case of doubt.
The output and the logfile in turn are used to
improve the filter programs in order to minimize
manual post-editing. This modular bootstrapping
approach has advantages over huge parameteriz-
able programs: filters are relatively simple and can
be partially reused or easily adapted for texts with
different formats; tuning the filters becomes less
complex; when recovering from a previous stage
the loss of work is minimized. The filters have
been implemented in Perl which, due to its pat-
tern matching mechanism via regular expressions,
is a very powerful language for such applications.
For the linguistic annotation we use the MUL-
TEXT tools available from v-
aix.fr/projects/multext. We already have exten-
sive experience with the tokenlzer MtSeg which

distinguishes 11 classes of tokens, such as abbrevi-
ations, dates, various punctuations, etc. The cus-
tomization of MtSeg via language-specific resource
files has been done in a bootstrapping process sim-
ilar to the filter programs. An evaluation of 10%
of the Civil Code (~ 28,000 words) revealed only
one type of tokenization error: a full stop that is
not part of an abbreviation and is followed by an
uppercase letter is recognized as end-of-sentence
marker, e.g. in "6. Absatz". This kind of error is
unavoidable in German if we refuse to mark such
patterns as compounds.
Currently we are preparing the lemmatization
and the POS tagging by using MtLex. MtLex is
equipped with an Italian and a German lexicon
which contain 138,823 and 51,010 different word
forms respectively. To include the 15,013 (58,217)
new Italian (German) word forms in our corpus
the corresponding lexicons have been extended.
The creation of the Italian lexicon took 2 MM.
Future work will include the completion of the
linguistic annotation. The MULTEXT tagger Mr-
Tag will be used for the disambiguation of POS
tags. Word alignment still requires the study
of various approaches, e.g. (Dagan et al., 1993;
Melamed, 1997). Finally, we are working on a so-
phisticated interface to navigate through parallel
documents to disseminate the text corpus before
terminology extraction has been completed.
References

Lynne Bowker. 1996. Towards a corpus-
based approach to terminography. Terminol-
ogy, 3(1):27-52.
Ido Dagan and Kenneth W. Church. 1997. Ter-
might: Coordinating humans and machines
in bilingual terminology acquisition. Machine
Translation, 12:89-107.
Ido Dagan, Kenneth W. Church, and William A.
Gale. 1993. Robust bilingual word alignment
for machine aided translation. In Proceedings
of the Workshop on Very Large Corpora: Aca-
demic and Industrial Perspectives, pages 1-8.
Nancy Ide, Greg Priest-Dorman, and Jean
V~ronis. 1996. Corpus encoding standard. See

I. Dan Melamed. 1997. A portable algorithm for
mapping bitext correspondence. In Proceedings
of ACL/EACL-97, pages 302-312.
276